AI writing tools still can't revise — researchers add a reflection mechanism to fix it

A research team found that current AI language models struggle with open-ended writing tasks because they lack the ability to reflect on and revise their own work, unlike their strong performance on math problems. They built a new system that adds explicit reflection and revision loops, allowing the AI to evaluate its own writing and iterate — producing measurably better results on creative writing and research tasks.

Most AI advances in reasoning have come from mathematical and verifiable domains where there's one right answer to check against. Writing and research tasks are fuzzier — there's no clear signal telling the AI whether something is good or needs rework. This paper shows that teaching AI to explicitly reflect and critique its own output, rather than just generating something once, unlocks capabilities that were previously invisible. It means the bottleneck wasn't the AI's raw ability — it was the training method.